Learning from Symmetry: Meta-Reinforcement Learning with Symmetrical Behaviors and Language Instructions